Query Fan-Out Explained: How AI Search Breaks One Question Into Many

Query fan-out is Google AI Mode splitting one question into many sub-queries, running them at the same time, and merging the results into one answer, which means your page is competing for a sub-question, not the head term you track. Write one clear answer per sub-question and you give the engine something to cite.

This applies to complex, multi-part questions, because simple factual questions trigger little or no fan-out.

Google has published the mechanism but not the sub-queries behind any given search, so every sub-query list, including the one in this guide, is a planning estimate.

What Query Fan-Out Means In Google AI Mode

Most of what you rank for was never one search. It was a pile of smaller questions the searcher never typed. AI Mode now goes and asks them for the searcher.

One note before the definition. "Fan-out" is also a term in hardware, networking and software engineering. This guide uses the AI search meaning only.

• Definition: Query fan-out is the set of concurrent, related queries that Google AI Mode generates from one user question, so the model can request more information than the original wording asks for. Google's developer guide defines it this way, and Google's own description is that AI Mode splits a question into subtopics and sends many searches at the same time.

Google's AI Mode announcement made that description public, and the update was announced on May 20, 2025, with AI Mode reaching all US users in June 2025.

Not every question gets this treatment. Simple factual queries trigger little or no fan-out, while complex queries trigger extensive fan-out. "What is the capital of France" needs one lookup. "How should a 50-person agency choose project management software" needs a dozen, and that second kind of question is where your content gets chosen or skipped.

If you want the wider concept of being mentioned and cited in AI answers, start with our guide to AI search visibility. The rest of this guide stays on one mechanism and one method.

How AI Mode Splits One Question Into Sub-Queries

The process runs in four steps. The order matters, because your page only gets a chance at step three and only gets cited at step four.

1. Decompose the question. The model reads the full question and works out which subtopics it contains.
2. Generate the sub-queries. Each subtopic becomes one or more searches, written by the model, not by the user.
3. Retrieve in parallel. The searches run at the same time against internal sources such as the Knowledge Graph, shopping data and news.
4. Synthesise one answer with citations. The model combines what came back into a single response and links the sources it decides to use.

The final answer does not show every underlying query. You see the response and its links, and the searches behind it stay out of view.

Retrieval is not citation. A page can come back from a sub-query and never be opened or linked in the answer. That is why this guide spends its time on how to make a passage worth choosing, not on how to get retrieved.

Two practical points follow from the mechanism. First, the engine is matching passages to sub-questions, so a page that answers one sub-question well beats a page that vaguely touches ten. Second, the sub-queries are written by a model, so they will not always match the phrasing any keyword tool gives you. You plan for the facet, not for the exact string.

How Many Sub-Queries Does One Question Produce?

Nobody outside Google can give you a typical number. No primary source states one, so plan for a range and treat every figure below as a data point, not a rule.

Google's blog post says Deep Search runs hundreds of searches for a single question. That is the top end of the range, and it is not what an everyday AI Mode query does.

For everyday queries, Ahrefs' analysis found that AI Mode typically runs 5 to 11 searches per query, while ChatGPT Deep Research made 420. The second figure measures a different product doing a different job, so do not read it as an AI Mode number.

One worked case shows what a single run looks like. In an example from Aleyda Solis, one headphone shopping prompt broke into 13 sub-queries covering comfort, battery, reviews, brands and price. It is a single example, not a rate.

So what do you plan around? A range of roughly five to a dozen sub-questions for a complex query is a sensible working assumption, and the next two sections show how to turn that into a list you can act on.

Types Of Fan-Out Queries

Sub-queries are not random. They fall into facets, and facets are what you map your content against. The headphone example gives the shopping version, and a general version covers most B2B topics.

The headphone prompt split into five facets.

• Comfort: Questions about how the product feels over long use. An example sub-query is "are these headphones comfortable for all-day wear".
• Battery: Questions about how long it lasts and how it charges. An example sub-query is "wireless headphones battery life per charge".
• Reviews: Questions about what owners and testers say. An example sub-query is "best rated wireless headphones reviews".
• Brands: Questions about which makers are worth considering. An example sub-query is "top headphone brands for commuting".
• Price: Questions about what it costs. An example sub-query is "wireless headphones under a set budget".

For software and services, a five-way grouping covers most of what a buyer asks.

• Definition: The sub-question that asks what something is. An example sub-query is "what is a project management tool".
• Process: The sub-question that asks how to do it. An example sub-query is "how to roll out project management software to a team".
• Comparison: The sub-question that asks how options differ. An example sub-query is "project management software for agencies compared".
• Cost: The sub-question that asks what it costs and what drives the price. An example sub-query is "project management software pricing per seat".
• Risk: The sub-question that asks what can go wrong. An example sub-query is "project management software adoption problems".

Keep these five as your default categories. They are the columns of the method later in this guide.

What Fan-Out Changes About Which Pages Get Cited

In classic search, one query returns one ranked list, and one page sits at the top of it. With fan-out in Google AI Mode and AI Overviews, one question becomes many sub-queries, and each sub-query has its own set of passages that could be cited.

The practical result is that a page can be cited without ranking for the head term. If your page answers one sub-question better than anything else retrieved for it, the engine can use that passage and ignore the rankings for the original wording.

AI search breaks pages into chunks, which are small, context-rich passages that answer precise questions, and each fan-out branch can become a standalone passage. Treat that as how the retrieval is generally described, not as a tested result. It still tells you where to put effort. A section that addresses a facet-level question gives the engine something to extract. A page that only covers the head term in general language gives it nothing to pick.

This is also different from long-tail keywords. A long-tail keyword is a search a person types, with volume you can look up. A fan-out query is a search the model writes, often with no volume figure at all, so a keyword tool will miss some of them.

Google's stance is plain. Its guidance says no special AI optimisation or AI-specific markup is required and that standard SEO still applies. That matches what we would expect, and it means the work here is content structure, not a new technical layer. For the practical steps on the AI Overviews side, read how to increase Google AI Overview visibility.

If you want a team to own this kind of work, our Answer Engine Optimization service is built around being the cited answer, and it starts from the sub-questions, not the head term.

Rankings still matter. What changes is that ranking first for the head term stops being the only route to a citation.

Does Query Fan-Out Work The Same In ChatGPT And Other Engines?

The idea is shared across AI engines. The evidence is not. Each engine decides for itself when to search and how many times, and the public data we have is thin.

The one ChatGPT figure we can state comes from an October 2025 study in which about 31% of 8,500+ commercially oriented ChatGPT prompts recorded web-search activity. That figure measures ChatGPT, not Google AI Mode, and it does not tell you how many sub-queries each search produced.

So the sensible reading is simple. Build content that answers sub-questions well and it should serve more than one engine, but do not carry an AI Mode number over to ChatGPT or the other way round. For how the labels for optimising across engines relate to each other, see our breakdown of AISO versus SEO, AEO and GEO.

How To Map One Target Query To Its Sub-Queries

This is the method. It is a planning method, not a process Google publishes, so the output is a working hypothesis you test, not a record of what Google ran.

A common approach is to map a root question into sub-questions, group them by facet, then check which ones your existing content leaves unanswered. Here it is in six steps.

1. Pick one target query. Choose a query that matters commercially and is complex enough to trigger fan-out. The output is a single query written on a blank page.
2. State the buyer question behind it. Write the real question in one sentence, including who is asking and what they need to decide. The output is one sentence you can test every sub-query against.
3. List the likely sub-queries by facet. Go through definition, process, comparison, cost and risk, and write two or three sub-queries under each. The output is a list of roughly five to fifteen sub-queries, which fits the range from the counts section.
4. Mark each sub-query as owned, thin or missing. Owned means a section on your site answers it directly. Thin means a section touches it but does not lead with an answer. Missing means nothing answers it. The output is a status column.
5. Assign a section or page to each one. For owned and thin items, name the section to fix. For missing items, name where the answer will live. The output is an action column.
6. Rank the gaps. Put missing items that sit close to the buying decision first, and thin items second. The output is a ranked gap list you can hand to a writer.

A few habits make the list better. Write sub-queries as a searcher would phrase them, not as headings. Include at least one sub-query that the buyer would be embarrassed to ask a salesperson, because buyers often have that question. And keep the list short enough that you will actually act on it.

Be honest about the limits. You cannot verify that Google ran any of these queries. What you can verify is whether your content answers the question, and that is the part you control.

Worked Example: Mapping A B2B SaaS Query

Take an invented but realistic query. The target is "how to choose project management software for a 50-person agency", and the buyer question behind it is which tool fits a services team that bills by project and juggles many clients.

The sub-queries below are illustrative and were not taken from Google. The status column shows how a real audit might look for a vendor with a typical product site.

Read the gaps with one rule. Sections close in-page gaps, and a new page is justified only when the sub-question needs its own depth.

Here is how that plays out in the table. The thin items become rewrites of existing sections, because the answer exists and needs to lead. The two missing items on the main guide become new sections, because the buyer reads them in the same sitting. The comparison for client-facing reporting is best answered somewhere other than your own site, because a buyer asking for the best option trusts third parties more than a vendor. Flag it as an off-site gap and plan for it there, as part of a wider approach.

How To Find The Sub-Queries You Cannot See

Google does not show the queries it runs. The final answer does not reveal every one, so you build a likely list from several angles and accept that it is an estimate.

Start with the answer itself. Run your target prompt in AI Mode and read the structure of the response. The headings, the order of the points and the comparisons it makes all hint at the facets the model worked through. Then add People Also Ask questions from the classic results, which often line up with the same facets.

For a first set of sub-question candidates, Semrush's Keyword Magic Tool can filter a topic down to question-style keywords, which gives you a list to check against your pages before you test anything in an AI engine.

Tools split into two groups by the job they do.

• Expansion tools: AlsoAsked, AnswerThePublic and Keyword Insights surface sub-questions and related questions for a topic.
• Testing and tracking tools: These check whether your passages are retrieved and whether your brand is cited when you run a prompt.

Use expansion tools to build the list and testing tools to see whether the list is working. If you are choosing a tracking platform, our AI visibility platform buyer guide compares what matters.

Do not forget off-site sources. Some sub-questions are answered by third-party pages the engine prefers, and source discovery tools show which domains AI engines cite for your topic. Those domains are your off-site gap list.

Finally, keyword research still has a role. Search volume will not tell you what the model asked, but it shows which facets real buyers care about, and that is a fair tiebreaker when your list is long.

How To Structure A Page So Each Sub-Question Gets Answered

Once you have the gaps, the fix is page structure. Each fan-out branch can become a standalone passage, so each section has to work when it is lifted out of the page.

Use this checklist for every section that answers a sub-question.

• Answer first: Open the section with the direct answer. The engine looks for a passage it can use whole, and a section that builds up to its point gives it nothing to lift.
• Name the subject: Write the entity in full instead of "it" or "this tool". A passage with no context around it has to carry its own subject.
• Two to four sentences: Kopp says in Digiday that passages should ideally run 2 to 4 sentences.
• One claim per passage: A passage that makes three claims can only be cited for one, and the other two dilute it.

Headings do real work here. Write one sub-question per section, and make the heading describe that sub-question in plain words.

• Before: Our Approach To Onboarding
• After: How Long Does It Take To Roll Out Project Software To A 50-Person Team?

The second heading tells a reader and an engine exactly what the section answers. If you want the technical side, including content models, schema and internal linking, read our piece on content engineers building for AI search.

Should You Build A Page For Every Sub-Query?

No, and Google has said why. Its guidance states that creating separate pages for each fan-out variation, mainly to manipulate rankings or AI responses, is a scaled content abuse violation. Ten thin pages for ten sub-queries is the exact pattern it warns against.

The default is a section on an existing page. Build a new page only when the sub-question has its own search demand and enough depth to deserve one. Page count alone does not earn AI citations, and whether one comprehensive page or a linked cluster works depends on the subject.

The off-site row matters more than it looks. Some sub-questions, like "which tool do other teams prefer", are credible only when someone other than you answers them. Treat those as gaps to flag, with the reasoning that a buyer trusts an independent answer more, and plan for mentions elsewhere.

If you want help deciding which sub-questions justify a page and which do not, our SaaS Content Marketing work starts from exactly that planning step.

How To Check Your Sub-Query Coverage Over Time

You cannot see the queries, but you can see the outcomes. Check coverage the same way each time, so the changes mean something.

1. Fix a prompt set. Write ten to twenty prompts based on your target queries and sub-queries, and keep the wording the same each run.
2. Run the set on a schedule. Run it in AI Mode and any other engine you care about, and record whether your brand is mentioned and whether your pages are cited. Our guide to AI search visibility metrics and KPIs defines these measures.
3. Compare against the gap list. Check each result against the owned, thin and missing statuses from your map. A thin item that now gets cited has moved, and a missing item that still gets nothing has not.
4. Redo the map when something changes. If the offer, the audience or the topic changes, or the same gaps stay uncited after you have fixed them, go back to the mapping step and rebuild the list.

Measure mentions and citations with a consistent prompt set, since no tactic guarantees citations. If you want a tool to run this for you, see our comparison of tools for tracking brand visibility in AI search.

For an example of measured topic coverage work, read our own AEO case study, which shows what we tracked on our own site and how.